Professor/yoruba-en-ner-model

Yoruba-English Code-Switching Language Identification (LID)

This model is a fine-tuned version of AfroXLM-R-Large designed to identify language boundaries in Yoruba-English code-switched text at the token level.

Model Description

The model classifies each token in a sentence into one of three categories:

• YORUBA: Tokens belonging to the Yoruba language.
• ENGLISH: Tokens belonging to the English language.

By utilizing the AfroXLM-R-Large backbone, which was pre-trained with a focus on African languages, this model demonstrates exceptional robustness in handling the morphological complexities of Yoruba and the fluid transitions in code-switched speech.

Performance (Test Set)

The model achieved near-perfect performance. Peak generalization was reached at Epoch 1. While training continued for 5 epochs for observation, the final deployed weights are from the first epoch to ensure maximum generalizability and prevent over-memorization of training samples.

Class	Precision	Recall	F1-Score	Support
Overall	0.991	0.991	0.991	~80k
English	0.995	0.994	0.994	63,016
Yoruba	0.976	0.979	0.978	17,069

Intended Uses & Limitations

Intended Use

• Research in Code-Switching (CS) patterns.
• Preprocessing for Machine Translation or Speech Synthesis (TTS) involving Yoruba-English bilingual speakers.
• Computational linguistics studies on the matrix language frame in Nigerian English.

Limitations

• Tonal Markers: Performance may slightly vary if Yoruba text lacks standard diacritics (tonal marks).
• Domain Sensitivity: Optimized for general conversational and science-related prompts; performance on archaic or highly legalistic Yoruba may vary.

Training Procedure

Hyperparameters

• Base Model: AfroXLM-R Large (550M parameters)
• Batch Size: 128 (Global)
• Learning Rate: 3e-05 (with Cosine Decay)
• Precision: BF16 (Brain Floating Point)
• Optimizer: AdamW (Fused)

Training Narrative

The model converges remarkably fast due to the pre-existing linguistic knowledge in the AfroXLM-R base. Users will notice that Validation Loss is lowest at Epoch 1.0 ($0.0240$). Despite the training loss continuing to drop, the validation loss begins a slight upward trend thereafter, indicating that the model captures the underlying linguistic boundaries almost immediately.

How to Use

from transformers import pipeline

lid_model = pipeline("token-classification", model="your-username/yoruba-en-ner-model")
text = "Egungun eleru helps to cleanse the village by carrying ebo"
results = lid_model(text)

for entity in results:
    print(f"Token: {entity['word']}, Language: {entity['entity']}")

Citation

If you use this model in your research, please cite the Masakhane AfroXLM-R paper and this fine-tuned version.